Date of Award

9-3-2019

Publication Type

Master Thesis

Degree Name

M.A.Sc.

Department

Electrical and Computer Engineering

Supervisor

Tepe, K.

Rights

info:eu-repo/semantics/openAccess

Abstract

Internet of Things (IoT)- based remote health monitoring systems have an enormous potential of becoming an integral part of the future medical system. In particular, these systems can play life-saving roles for treating or monitoring patients with critical health issues. On the other hand, it can also reduce pressure on the health-care system by reducing unnecessary hospital visits of patients. Any health care monitoring system must be free from erroneous data, which may arise because of instrument failure or communication errors. In this thesis, machine-learning techniques are implemented to detect reliability and accuracy of data obtained by the IoT-based remote health monitoring. A system is a set-up where vital health signs, namely, blood pressure, respiratory rate, and pulse rate, are collected by using Spire Stone and iHealth Sense devices. This data is then sent to the intermediate device and then to the cloud. In this system, it is assumed that the channel for transmission of data (vital signs) from users to cloud server is error-free. Afterward, the information is extracted from the cloud, and two machine learning techniques, i.e., Support Vector Machines and K-Nearest Neighbor are applied to compare their accuracy in distinguishing correct and erroneous data. The thesis undertakes two different approaches of erroneous data detection. In the first approach, an unsupervised classifier called Auto Encoder (AE) is used for labeling data by using the latent features. Then the labeled data from AE is used as ground truth for comparing the accuracy of supervised learning models. In the second approach, the raw data is labeled based on the correlation between various features. The accuracy comparison is performed between strongly correlated features and weakly correlated features. Finally, the accuracy comparison between two approaches is performed to check which method is performing better for detecting erroneous data for the given dataset.

Share

COinS